Chapter IV: UCNA: Fierz Interference Analysis
4.2 The Spectral Extraction Method
4.2.1 The Super-Sum Spectrum
4.2.1.3 Blinding the Spectral Extraction
After the publication of the UCNA 2010 dataset Fierz interference direct spectral extraction (summarized later in section 4.2.5), there was a move towards blinding the data for the subsequent UCNA 2011-2012 and 2012-2013 datasets to ensure a more robust analysis. For an in-depth discussion on blinding data analyses in experimental physics, see [KR05]. In summary, there are several case studies of physics experiments throughout history that show unconscious bias ultimately leading to convergences in experimental measurements. These convergences are non-statistical in nature and their prevalence can be explained by unconscious bias.
Blinding thus was introduced as a method of combating this human factor and has lead to a successful reinterpretation of experimental measurements, producing a distribution that fluctuates within experimental uncertainties. The community as a whole has generally concluded blinding to be a crucial step in validating the integrity of experimental analysis results. With this in mind, we decided to blind the 2011-2012 and 2012-2013 Fierz interference extractions. Below, we describe the blinding methodology for the direct spectral analysis.
Recall from the previous section, section 4.2.1.2, that we use two baseline Monte
Carlo histograms, processed through the GEANT4 simulation and with the detector response model to become βdata-likeβ. We then use these data-like histograms to fit the real data histogram and extract a fractional presence of the π =0 histogram and theπ =βhistogram. These fractional percentages directly relate to aπ value.
In order to perform the blinding, there would be no way to reliably modify the data since we only have one true dataset and a priori we do not know whatπ value that dataset would extract. Instead, what we do is modify the baseline Monte Carlo histograms. By βmixing inβ a percentage of events from theπβ βhistogram, we distort the π = 0 histogram and transform it into a π β 0. In particular, when we mix inπβ βevents into theπ =0 histogram, the baselineπ=0 histogram shifts toπ > 0 (the energy peak shifts to lower energy, which physically is what a π > 0 distortion induces) and hence when we fit a standard π = 0 histogram with these two βblindedβ baseline histograms, we would produce aπ < 0 value. We note this is the first βcaseβ: creating a blindedπ < 0 value.
To show this explicitly, we start by taking the results of equation 4.5, ππ(πΈπ) =
1+π(ππ
πΈπ) 1+π
Dππ πΈπ
Eππ π(πΈπ), and that in equation 4.6, ππββ = ππ πππ
πΈπ
Dππ
πΈπ
Eβ1
. Also, the fitting is defined in equation 4.7 as ππππ‘ π(πΈπ) =
ππ π(πΈπ)+π Dππ
πΈπ
E
ππββ(πΈπ) 1+π
Dππ πΈπ
E .
We now introduce a mixing percentage,π, to our π = 0 distribution. That is, each event in theπ =0 histogram is sampled and with probabilityπ that event is replaced by the corresponding numbered event from the π β βhistogram. We note both generated histograms have the same number of events. This results in the following transformation on our baseline SM distribution:
ππ π β¦β (1βπ)ππ π +π ππββ (4.8) where ππ π and ππββ have implicit energy dependence that we suppress for the sake of notation here. Replacingππ π in equation 4.7 with the result from equation 4.8, we get
ππππ‘ π(πΈπ) = π0
π π(πΈπ) +πhππ
πΈπ
iππββ(πΈπ)
1+πhππ
πΈπ
i
=
(1βπ)ππ π(πΈπ) +π ππββ(πΈπ) +πhππ
πΈπ
iππββ(πΈπ)
1+πhππ
πΈπ
i
=
(1βπ)ππ π(πΈπ) + [π+πhππ
πΈπ
i]ππββ(πΈπ)
1+πhππ
πΈπi
(4.9)
The goal here is to introduce a falseπvalue that then becomes our blindedπvalue.
In order to determine theπ value introduced by a mixing percentageπ, we replace the ππππ‘ π(πΈπ)distribution with aπ=0 Monte Carlo distribution. This gives us
ππ π(πΈπ) =
(1βπ)ππ π(πΈπ) + [π +πhππ
πΈπ
i]ππββ(πΈπ)
1+πhππ
πΈπ
i (4.10)
For equation 4.10 to be true, we can set the coefficient in front ofππ π to 1 and the coefficient in front ofππββ(πΈπ)to 0. Solving both of these gives the same answer since this problem is over-constrained.
β΄ π= βπ hππ
πΈπ
i (4.11)
π is the Fierz interference value we are trying to introduce as a blinding and π is a percentage of events we are mixing between a SM distribution and a π β β distribution.
The second case is producing a blinded π > 0 value. We can repeat the same procedure except we mixπ =0 andπ =β1 events together. Due to the form ofπin equation 3.1, π =β1 is the lowest possible value (any more negative values would induce a negative decay rate in the equation). Henceπ =β1 conceptually represents the opposite of π β β, an effective π β ββ. By mixing inπ = β1 events, the baseline Monte Carlo histogram transform from π = 0 to π < 0 (the energy peak shifts to a high energy) and hence when we fit a standard π = 0 histogram, we produce aπ > 0 value.
To show this explicitly, we again use a mixing percentage, π, of π = β1 events in ourπ =0 distribution. This results in the following transformation on our baseline SM distribution:
ππ π β¦β (1βπ)ππ π +π ππ=β
1 (4.12)
whereππ π andππ=β
1(andππββbelow) have implicit energy dependence that we again suppress for the sake of notation. When we use equations 4.5 and 4.6, we also get
ππ=β
1= 1β ππ
πΈπ
1β hππ
πΈπ
iππ π
= ππ π 1β hππ
πΈπ
i β
ππ
πΈπ
ππ π 1β hππ
πΈπ
i
= 1 1β hππ
πΈπi
ππ πβ hππ πΈπ
iππββ
(4.13)
Plugging this result into equation 4.12, the new baseline distribution we are fitting with becomesππββand
ππ π β¦β (1βπ)ππ π+ π 1β hππ
πΈπi
ππ πβ hππ πΈπ
iππββ
(4.14)
Now we collect the terms in front of ππ π and ππββ and equate this to a π = 0 Standard Model histogram, giving
ππ π =
(1βπ) + π
1βhπ
πΈi
1+ hπ
πΈi ππ π +
πhπ
πΈi + πh
π πΈi 1βhπ
πΈi
1+ hπ
πΈi ππββ (4.15) In order for equation 4.15 to hold, the coefficient in front ofππ π must be 1 and the coefficient in front ofππββmust be 0. Again, solving both of these gives the same answer since this problem is over-constrained.
β΄ π= π 1β hπ
πΈi (4.16)
where, again,π is the Fierz interference term we are trying to introduce andπ is a percentage of events we are mixing between a π = 0 (SM) spectrum and aπ = β1 (maximal Fierz in positive energy shift for the spectrum peak) spectrum.
Equations 4.11 and 4.16 represent the blinded value of π chosen by sampling a percentage of events from aπ β βdistribution and aπ=β1 distribution and using them to replace the same number of events from aπ = 0 distribution. Afterwards, we tested equations 4.11 and 4.16 by fitting a known inputπvalue with equation 4.7 and confirmed that they produced a blindedπvalue when fitting aπ =0 spectrum.
We then randomly sampled which formula to use, corresponding to whether we blinded withπ < 0 orπ > 0, and then chose an unknown value ofπ corresponding to a blinding valueπ β [β0.075,0.075].